Artificial neural networks can organize their own learning machinery, MIT study finds

MIT researchers report in PNAS that artificial neural networks can build their own learning machinery when given simple rules for how connections change.

Randomly connected units with a "neurons that fire together wire together" rule self-organized into motifs that classified images, performing at a level similar to networks trained with back-propagation.

The authors argue that self-organization is a biologically plausible measure of a learning model, since the brain has no external designer.